cond-mat.dis-nnMar 16, 2026

Machine learning cannot decarbonise the subsurface until it can be audited

Authors: Hannah P. Menke, Ahmed H. Elsheikh, Lingli Wei, Nanzhe Wang, Andreas Busch

Organizations: Institute of GeoEnergy Engineering, Heriot-Watt University, Edinburgh, UK · Subsurface Energy Transition and Innovation Centre, Heriot-Watt University, Edinburgh, UK · SubsurfaceLabs, Abu Dhabi, UAE · Lyell Centre, Heriot-Watt University, Edinburgh, UK

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Geological CO2 storage, geothermal energy and underground hydrogen storage depend on irreversible or high-cost decisions made with sparse data and incomplete physical understanding. Machine learning (ML) increasingly informs these decisions, yet methods are usually judged on predictive accuracy rather than on whether their outputs can be trusted, traced and acted upon. We argue that subsurface ML should be designed as decision infrastructure with shared systems that carry calibrated uncertainty and provenance from observation to operating decision. We propose a decision-readiness checklist, an audit-trail workflow for monitoring updates, and benchmark designs that score models by the decisions they support. We identify validation needs shared across the three technologies, and the data-access and governance steps needed for ML to support regulatory and investment decisions.

Figures & tables

Figure 1: Funnel ×\times Stack framing: the subsurface project funnel (basin screening →\rightarrow prospecting →\rightarrow exploration and appraisal →\rightarrow operations and MMV →\rightarrow closure and stewardship) repeated through the ML stack (sensing →\rightarrow inference →\rightarrow simulation →\rightarrow decision). The uncertainty ribbon highlights that uncertainty is generated at early funnel stages, shaped by new data and models as the project matures, and must ultimately be expressed as decision risk (operating envelopes and trigger criteria). THCM, thermo-hydro-chemical-mechanical.
 CO 2 storageGeothermal / EGSH 2 storage
Shared data typesSeismic surveys, well logs, core and SCAL data, downhole pressure and temperature, microseismic monitoring, legacy reports and analogues.
Shared ML methodsUncertainty-aware surrogates, ensemble data assimilation, structure-aware geological priors, hybrid physics-ML, multi-fidelity learning, value-of-information monitoring design.
Primary decision variablesContainment probability, pressure footprint, plume conformanceThermal breakthrough time, induced seismicity risk, sustainable flow rateCyclic deliverability, recovery efficiency, containment under cycling
Dominant coupled physicsMultiphase flow, dissolution and residual trapping, pressure-driven geomechanicsHeat transport, fracture-dominated thermo-hydro-mechanical couplingHysteretic multiphase flow under cycling, microbial and geochemical H 2 reactions, cushion-gas mixing
Operational regimeOne-way injection over decades; stewardship over centuriesContinuous circulation over years to decadesSeasonal to weekly injection-withdrawal cycles
System-specific dataTime-lapse (4D) seismic, surface deformation from interferometric synthetic aperture radar (InSAR), geochemical assurance monitoringTracer tests, temperature and flow logs, dense microseismic cataloguesPer-cycle pressure and gas-composition histories, well deliverability tests
Table 1: Shared structure and system-specific differences across the three subsurface decarbonisation technologies. Rows above the midline are largely shared; rows below differ materially and require system-specific treatment.
DiagnosticWhat it measuresDecision consequence
Reliability diagram / calibration curveWhether stated probabilities match observed frequenciesSupports traffic-light decisions when calibration is verified at the relevant risk levels and under the intended operating conditions
Prediction-interval coverageFraction of held-out outcomes falling inside stated intervalsIllustratively, 90% intervals covering 60% of held-out pressure data indicate undercoverage; recalibrate uncertainty and reassess pressure-limit exceedance before operational use
Continuous ranked probability score (CRPS)Sharpness and calibration of the full predictive distribution 22Ranks competing forecast systems for pressure or plume evolution on a proper basis
Rank / probability-integral-transform histogramsEmpirical rank or PIT distributions relative to uniformityAssess calibration across operating regimes and evaluate tail-event probabilities directly for seismicity and leakage decisions
Expected calibration error (classifiers)Aggregate miscalibration of event probabilitiesSummarises probability calibration; false-alarm and missed-detection rates require threshold-specific evaluation
Conformal coverageDistribution-free finite-sample marginal coverage under exchangeability 24Requires decision-specific validation and validity-domain monitoring; distribution shift can invalidate coverage
Table 2: Calibration diagnostics for decision-ready uncertainty, and the decision consequence of each.
Table 4
Figure 2: Illustrative decision output for subsurface decarbonisation: an operating envelope / Pareto front showing trade-offs under uncertainty (e.g. performance versus pressure footprint and induced seismicity risk). The deliverable is a calibrated, risk-bounded frontier.
Table 6
FieldMature practiceTransfers to subsurfaceDoes not transfer
Numerical weather predictionContinuous ensemble verification; proper scoring rules (CRPS); routine reliability reporting 22 ; 23Scoring MMV and deliverability forecasts like weather forecasts; calibration as a reported normVerification volume: millions of daily outcomes versus a handful of slow, confounded field outcomes
Clinical MLTRIPOD+AI reporting standard 65 ; model cards 66 ; device approval pathwaysStructured reporting of data, validation, intended use, and failure modes; pre-deployment reviewPatient-scale cohorts; randomised evaluation of the decision itself
Aviation and nuclear safetyMandatory incident reporting; shared failure-mode catalogues; safety casesStandardised failure-mode reporting and safety-case structure for ML workflowsComponent-level failure statistics built from large fleets
Table 3: What transfers from adjacent safety-critical fields, and what does not.
Table 4: Minimum viable ecosystem requirements, organised by actor, as an illustrative near-term (two-to-three-year) horizon.
More in this topic

Scientific ML

Scientific machine learning combines data-driven models with scientific knowledge to analyze, predict, and simulate complex systems.

See all 122 papers

Explore similar work

CardsList

  1. Physics-Aware Machine Learning for Seismic and Volcanic Signal Interpretation

    Mar 18, 2026William ThorossianPhysics-Informed MLSeismology

  2. Towards Fast GNN Surrogates for CO2 Migration in Complex Geological Formations

    Jun 15, 2026Rodrigo S. Luna, Thiago H. N. Coelho, Luiz S. L. Neto +8Graph Neural NetworksPDE Surrogate Modeling

    Systems and Computer Engineering and High Performance Computing Center, NACAD - COPPE, Federal University of Rio de Janeiro · Civil Engineering and High Performance Computing Center, NACAD - COPPE, Federal University of Rio de Janeiro · Mechanical Engineering and High Performance Computing Center, NACAD - COPPE, Federal University of Rio de Janeiro +1

  3. GeoCFM: Positive-Only Conditional Flow Matching for Mineral Occurrence Sampling

    Sep 28, 2026Moshe Eliasof, Eldad HaberConditional Flow MatchingGeospatial Reasoning

    Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Israel · Department of Earth, Ocean and Atmospheric Sciences, University of British Columbia, Canada